I merged your last two PRs and pushed an updated branch. The test fails now due to an operation like transaction(write 0~5, write 10~5) where the second write wants to read the block in order to write a full block/chunk, but the first write's data obviously isn't off to disk yet. The missing piece is a buffer cache, which we want anyway so that we can - handle cases like this - drop the hacky tail_cache code - cache data in ram (when we have an fadvise hint, or for some other reason decide it's a good idea) I have a really basic initial version pushed that just handles the write path (buffers initalling in writing state, then clean once io hits device). We just need to make _do_read use it (Igor?) and add some basic trimming. (And eventually decide how we want to do trimming in general.. presumably we want a global LRU instead of the simple per-object thing that's there now.) Once that is working, I think the write path will be pretty complete, and we can turn our attention to the compression/checksum decisions (when to compress, how big should the chunks be, etc.) and performance. See this PR: https://github.com/ceph/ceph/pull/6208 I also coded up a new _clone that'll do COW, but I haven't tested it yet. sage -- To unsubscribe from this list: send the line "unsubscribe ceph-devel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
